On 2021-05-24 19:15, Jakub Kicinski wrote:

On Mon, 24 May 2021 15:12:20 +0300 Maxim Mikityanskiy wrote:

quoted

@@ -1290,6 +1304,26 @@ static int tls_device_down(struct net_device *netdev)
  	spin_unlock_irqrestore(&tls_device_lock, flags);
  
  	list_for_each_entry_safe(ctx, tmp, &list, list)	{
+		/* Stop offloaded TX and switch to the fallback.
+		 * tls_is_sk_tx_device_offloaded will return false.
+		 */
+		WRITE_ONCE(ctx->sk->sk_validate_xmit_skb, tls_validate_xmit_skb_sw);
+
+		/* Stop the RX and TX resync.
+		 * tls_dev_resync must not be called after tls_dev_del.
+		 */
+		WRITE_ONCE(ctx->netdev, NULL);
+
+		/* Start skipping the RX resync logic completely. */
+		set_bit(TLS_RX_DEV_DEGRADED, &ctx->flags);
+
+		/* Sync with inflight packets. After this point:
+		 * TX: no non-encrypted packets will be passed to the driver.
+		 * RX: resync requests from the driver will be ignored.
+		 */
+		synchronize_net();
+
+		/* Release the offload context on the driver side. */
  		if (ctx->tx_conf == TLS_HW)
  			netdev->tlsdev_ops->tls_dev_del(netdev, ctx,
  							TLS_OFFLOAD_CTX_DIR_TX);

Can we have the Rx resync take the device_offload_lock for read instead?
Like Tx already does?

I believe you previously made this attempt in commit 38030d7cb779 
("net/tls: avoid NULL-deref on resync during device removal"), and this 
approach turned out to be problematic, as explained in commit 
e52972c11d6b ("net/tls: replace the sleeping lock around RX resync with 
a bit lock"): "RX resync may get called from soft IRQ, so we can't use 
the rwsem".

quoted

+EXPORT_SYMBOL_GPL(tls_validate_xmit_skb_sw);

Why the export?

Because tls_validate_xmit_skb was also exported. Now I see it's needed 
for tls_validate_xmit_skb, because tls_is_sk_tx_device_offloaded needs 
its address and can be called from the drivers. There is no similar 
public function for tls_validate_xmit_skb_sw, so you are probably right 
that we don't need to export it.